Re: HTB and ingress scheduler SMP Soft Lockup in 2.6.23-2.6.25-rc8

[Jarek Poplawski <jarkao2@gmail.com>]

Enrico Demarin wrote, On 04/11/2008 08:37 PM:

> Hi everyone,
> 
> I stumbled by chance in a deadlock condition using HTB and the ingress 
> scheduler at the same time.
> 
> - The bug seems to have been introduced in 2.6.23 and forward ( 2.6.22 
> works fine )
> - Tested up to  2.6.25-rc8
> 
> I found a temporary workaround  ( the ingress didnt have a reason to be 
> there in first place , just omitting it from the script avoids the 
> condition ), but it would be nice to have a proper fix. Let me know if 
> more data/testing is needed.

Hi,

It's a very nice test and results, but alas I don't have 3 boxes
around, so I wonder if you could try this debugging patch below?

I think one of possible reasons could be a qdisc with the same
handle and parent ids, like ingress, but I can't see how it could
be referenced by your htb classes yet...

Thanks,
Jarek P.

---

 net/sched/sch_api.c |    7 +++++++
 1 files changed, 7 insertions(+), 0 deletions(-)

diff --git a/net/sched/sch_api.c b/net/sched/sch_api.c
index 15b91a9..de52b17 100644
--- a/net/sched/sch_api.c
+++ b/net/sched/sch_api.c
@@ -386,11 +386,18 @@ void qdisc_tree_decrease_qlen(struct Qdisc *sch, unsigned int n)
 	if (n == 0)
 		return;
 	while ((parentid = sch->parent)) {
+		struct Qdisc *sch_old = sch;
+
 		sch = qdisc_lookup(sch->dev, TC_H_MAJ(parentid));
 		if (sch == NULL) {
 			WARN_ON(parentid != TC_H_ROOT);
 			return;
 		}
+		if (sch == sch_old) {
+			printk("qdisc loop: %p, %x\n", sch, sch->parent);
+			return;
+		}
+
 		cops = sch->ops->cl_ops;
 		if (cops->qlen_notify) {
 			cl = cops->get(sch, parentid);